U.S. patent application number 16/133803 was filed with the patent office on 2019-05-02 for information processing method, information processing apparatus, and program.
The applicant listed for this patent is Panasonic Intellectual Property Management Co., Ltd.. Invention is credited to MITSURU ENDO, ASUKA SAKAI, HONGJIE SHI, TAKASHI USHIO.
Application Number | 20190129932 16/133803 |
Document ID | / |
Family ID | 66243920 |
Filed Date | 2019-05-02 |
View All Diagrams
United States Patent
Application |
20190129932 |
Kind Code |
A1 |
SAKAI; ASUKA ; et
al. |
May 2, 2019 |
INFORMATION PROCESSING METHOD, INFORMATION PROCESSING APPARATUS,
AND PROGRAM
Abstract
An information processing method includes acquiring first text
information from a storage apparatus in which the first text
information representing one or more utterance sentences is stored
as a learning data set, identifying one or more named entities
included in the acquired first text information, replacing each of
the one or more identified named entities with an abstract
expression abstracted based on a predetermined rule thereby
generating second text information from the first text information,
and learning a model of a dialogue system using, as training data,
the second text information generated in the replacing.
Inventors: |
SAKAI; ASUKA; (Tokyo,
JP) ; ENDO; MITSURU; (Tokyo, JP) ; USHIO;
TAKASHI; (Tokyo, JP) ; SHI; HONGJIE; (Tokyo,
JP) |
|
Applicant: |
Name |
City |
State |
Country |
Type |
Panasonic Intellectual Property Management Co., Ltd. |
Osaka |
|
JP |
|
|
Family ID: |
66243920 |
Appl. No.: |
16/133803 |
Filed: |
September 18, 2018 |
Related U.S. Patent Documents
|
|
|
|
|
|
Application
Number |
Filing Date |
Patent Number |
|
|
62578740 |
Oct 30, 2017 |
|
|
|
Current U.S.
Class: |
1/1 |
Current CPC
Class: |
G06N 20/00 20190101;
G06F 40/295 20200101 |
International
Class: |
G06F 17/27 20060101
G06F017/27; G06N 99/00 20060101 G06N099/00 |
Foreign Application Data
Date |
Code |
Application Number |
Apr 27, 2018 |
JP |
2018-087391 |
Claims
1. An information processing method comprising: acquiring first
text information from a storage apparatus in which the first text
information representing one or more utterance sentences is stored
as a learning data set; identifying one or more named entities
included in the acquired first text information; replacing each of
the one or more identified named entities with an abstract
expression abstracted based on a predetermined rule thereby
generating second text information from the first text information;
and learning a model of a dialogue system using, as training data,
the second text information generated in the replacing.
2. The information processing method according to claim 1, wherein
in the replacing, in addition to replacing each of the one or more
identified named entities with the abstract expression abstracted
based on the predetermined rule, a numeral value is added depending
on a context in one or more utterance sentences represented in the
first text information thereby generating the second text
information from the first text information.
3. The information processing method according to claim 2, wherein
the numeral value depending on the context is a numeral value
depending on the order in which the one or more named entities
occur in the one or more utterance sentences represented in the
first text information.
4. The information processing method according to claim 2, wherein
in the replacing, a named entity, which is included in the one or
more identified named entities and which is of a facility name
indicating a target candidate that give a solution of the specific
task, is replaced with an abstract expression abstracted based on
the predetermined rule, and a numeral value corresponding to a rank
of an evaluation value about a service provided by a facility
indicated by the facility name is added as the numeral value
depending on the context.
5. The information processing method according to claim 1, further
comprising performing a process to solve a specific task using the
model learned in the learning, wherein in the acquiring, the third
text information is acquired from a storage apparatus in which the
third text information representing one or more utterance sentences
is stored as a test data set, in the identifying, one or more named
entities included in the acquired third text information are
identified, in the replacing, each of the one or more identified
named entities included in the third text information is replaced
with an abstract expression abstracted based on the predetermined
rule thereby generating fourth text information from the third text
information, and in the performing the process, the process to
solve the specific task is performed using, as a test data set, the
fourth text information generated in the replacing.
6. The information processing method according to claim 5, further
comprising performing preprocessing in which a bigram process is
performed as the prepressing to divide the second text information
and the fourth text information generated in the replacing into a
plurality of segments each including two adjacent expressions, in
the learning, the model is learned using, as training data, the
second text information subjected to the bigram process in the
preprocessing, and in the performing the process, a process is
performed to solve the specific task using, as a test data set, the
fourth text information subjected to the bigram process in the
preprocessing.
7. The information processing method according to claim 5, further
comprising performing preprocessing in which a standardization
process is performed as the prepressing to standardize a format of
the expressions included in the second text information and the
fourth text information generated in the replacing, wherein in the
learning, the model is learned using, as training data, the second
text information subjected to the standardization process in the
preprocessing, and in the performing the process, the process of
solving the specific task is performed using, as a test data set,
the fourth text information subjected to the standardization
process in the preprocessing.
8. The information processing method according to claim 5, further
comprising performing outputting such that one or more abstract
expressions, included in a result of the performed process to solve
the specific task and replaced based on the predetermined rule, are
replaced with the original named entities and output.
9. An information processing apparatus comprising: an acquirer that
acquires first text information from a storage apparatus in which
first text information representing one or more utterance sentences
is stored as a learning data set; an identifier that identifies one
or more named entities included in the acquired first text
information; a replacement processor that replaces each of the one
or more named entities identified by the identifier with an
abstract expression abstracted based on a predetermined rule
thereby generating second text information from the first text
information; and a learning processor that learns a model of a
dialogue system using, as training data, the second text
information generated by the replacement processor.
10. A program causing a computer to execute the information
processing method according to claim 1.
Description
BACKGROUND
1. Technical Field
[0001] The present disclosure relates to an information processing
method, an information processing apparatus, and a program, applied
to a dialogue system.
2. Description of the Related Art
[0002] A dialogue system is a system that executes, using a
dialogue robot or the like, a task depending on an utterance by a
user. There are two types of dialogue systems: a task-oriented
dialogue system that has a dialogue with a user to achieve a
specific task such as a restaurant search task; and a
non-task-oriented dialogue system that has a dialogue with a user
about unspecified various subjects.
[0003] The dialogue system needs to be capable of accurately
understanding a content of an utterance made by a user and properly
responding to the utterance. To achieve this requirement, it is
known to generate a model of the dialogue system by performing
learning such as machine learning or the like (see, for example,
Japanese Unexamined Patent Application Publication No.
2015-1910).
[0004] In a technique disclosed in Japanese Unexamined Patent
Application Publication No. 2015-1910, learning is performed to
make it possible to accurately determine a type of a dialogue
action even for an utterance including various informal
expressions.
SUMMARY
[0005] Utterance sentences used as training data can include a
named entity such as a proper noun or the like, which makes it
difficult to generate an accurate model of a dialogue system. In
the technique disclosed in Japanese Unexamined Patent Application
Publication No. 2015-1910, any particular consideration is taken on
learning an utterance sentence including a named entity.
[0006] In one general aspect, the techniques disclosed here feature
an information processing method including acquiring first text
information from a storage apparatus in which the first text
information representing one or more utterance sentences is stored
as a learning data set, identifying one or more named entities
included in the acquired first text information, replacing each of
the one or more identified named entities with an abstract
expression abstracted based on a predetermined rule thereby
generating second text information from the first text information,
and learning a model of a dialogue system using, as training data,
the second text information generated in the replacing.
[0007] The information processing method and the related techniques
according to the present disclosure makes it possible to achieve an
improvement in accuracy of the model of the dialogue system
generated via learning.
[0008] These general and specific aspects may be implemented using
a computer program, and any combination of systems, methods, and
computer programs.
[0009] Additional benefits and advantages of the disclosed
embodiments will become apparent from the specification and
drawings. The benefits and/or advantages may be individually
obtained by the various embodiments and features of the
specification and drawings, which need not all be provided in order
to obtain one or more of such benefits and/or advantages.
BRIEF DESCRIPTION OF THE DRAWINGS
[0010] FIG. 1 is a block diagram illustrating an example of a
configuration of an information processing apparatus according to
an embodiment;
[0011] FIG. 2 is a block diagram illustrating an example of a
detailed configuration of a data processor shown in FIG. 1;
[0012] FIG. 3 is a diagram illustrating an example of knowledge
stored in a knowledge base shown in FIG. 2;
[0013] FIG. 4 is a diagram illustrating examples of expression
pairs stored in a context dictionary shown in FIG. 2;
[0014] FIG. 5 is flow chart illustrating an outline of an operation
of an information processing apparatus according to an
embodiment;
[0015] FIG. 6A is a flow chart illustrating a detailed operation of
the information processing apparatus according to the
embodiment;
[0016] FIG. 6B is a flow chart illustrating another embodiment of a
process in step S103 shown in FIG. 6A;
[0017] FIG. 7 is a diagram illustrating an example of one dialogue
which is not yet subjected to a replacement process according to a
first embodiment;
[0018] FIG. 8 is a diagram illustrating an example of a context
dictionary according to the first embodiment;
[0019] FIG. 9 is a diagram illustrating an example of one dialogue
which has been subjected to the replacement process according to
the first embodiment;
[0020] FIG. 10 is a diagram illustrating an example of one dialogue
which is not yet subjected to a replacement process according to a
second embodiment;
[0021] FIG. 11 is a diagram illustrating an example of a context
dictionary according to the second embodiment;
[0022] FIG. 12 is a diagram illustrating an example of one dialogue
which has been subjected to the replacement process according to
the second embodiment;
[0023] FIG. 13 is a diagram illustrating an example of one dialogue
which is not yet subjected to a replacement process according to a
third embodiment;
[0024] FIG. 14 is a diagram illustrating an example of a context
dictionary according to the third embodiment;
[0025] FIG. 15 is a diagram illustrating an example of one dialogue
which has been subjected to the replacement process according to
the third embodiment;
[0026] FIG. 16 is a diagram illustrating an example of one dialogue
which is not yet subjected to a replacement process according to a
fourth embodiment;
[0027] FIG. 17 is a diagram illustrating an example of a context
dictionary according to the fourth embodiment;
[0028] FIG. 18 is a diagram illustrating an example of one dialogue
which has been subjected to the replacement process according to
the fourth embodiment;
[0029] FIG. 19 is a diagram illustrating an example of one dialogue
which is not yet subjected to a replacement process according to a
fifth embodiment;
[0030] FIG. 20 is a diagram illustrating a part of a knowledge base
according to the fifth embodiment;
[0031] FIG. 21 is a diagram illustrating an example of a context
dictionary according to the fifth embodiment;
[0032] FIG. 22 is a diagram illustrating an example of one dialogue
which has been subjected to the replacement process according to
the fifth embodiment;
[0033] FIG. 23 is a diagram illustrating an example of one dialogue
which is not yet subjected to a replacement process according to a
sixth embodiment;
[0034] FIG. 24 is a diagram illustrating a part of a knowledge base
according to the sixth embodiment;
[0035] FIG. 25 is a diagram illustrating an example of a context
dictionary according to the sixth embodiment;
[0036] FIG. 26 is a diagram illustrating an example of one dialogue
which has been subjected to the replacement process according to
the sixth embodiment;
[0037] FIG. 27 is a diagram illustrating an example of one dialogue
which is not yet subjected to a replacement process according to a
seventh embodiment;
[0038] FIG. 28 is a diagram illustrating a part of a knowledge base
according to the seventh embodiment;
[0039] FIG. 29 is a diagram illustrating an example of a context
dictionary according to the seventh embodiment;
[0040] FIG. 30 is a diagram illustrating an example of one dialogue
which has been subjected to the replacement process according to
the seventh embodiment;
[0041] FIG. 31 is a diagram illustrating an effect of a replacement
process according to an embodiment of the present disclosure;
[0042] FIG. 32 is a diagram illustrating an effect of a replacement
process according to an embodiment of the present disclosure;
and
[0043] FIG. 33 is a diagram illustrating an effect of a replacement
process according to an embodiment of the present disclosure.
DETAILED DESCRIPTION
[0044] In an aspect of the present disclosure, there is provided an
information processing method including acquiring first text
information from a storage apparatus in which the first text
information representing one or more utterance sentences is stored
as a learning data set, identifying one or more named entities
included in the acquired first text information, replacing each of
the one or more identified named entities with an abstract
expression abstracted based on a predetermined rule thereby
generating second text information from the first text information,
and learning a model of a dialogue system using, as training data,
the second text information generated in the replacing.
[0045] According to the present aspect, it is possible to learn the
model of the dialogue system using training data obtained by
replacing named entities included in utterance sentences with
abstract expressions in which variations of named entities are
eliminated based on the predetermined rule. Thus, an improvement in
the accuracy of the model of the dialogue system generated by the
learning is achieved.
[0046] For example, in the replacement process, in addition to
replacing each of the one or more identified named entities with
the abstract expression abstracted based on the predetermined rule,
a numeral value may be added depending on a context in one or more
utterance sentences represented in the first text information
thereby generating the second text information from the first text
information.
[0047] For example, the numeral value depending on the context may
be a numeral value depending on the order in which the one or more
named entities occur in the one or more utterance sentences
represented in the first text information.
[0048] In these aspects, as a result of adding a numeral value,
depending on a context, to each abstract expression obtained by the
replacement according to the predetermined rule, it becomes
possible to provide training data in which different entities of
abstract expressions belong to the same concept are distinguished
from each other in one dialogue, while they are regarded identical
to each other between different dialogues. This makes it possible
to achieve a further improvement in the accuracy of the model of
the dialogue system generated by the learning.
[0049] For example, in the replacement process, a named entity,
which is included in the one or more identified named entities and
which is of a facility name indicating a target candidate that may
give a solution of the specific task, may be replaced with an
abstract expression abstracted based on the predetermined rule, and
a numeral value corresponding to a rank based on an evaluation
value about a service provided by a facility indicated by the
facility name may be added as the numeral value depending on the
context.
[0050] For example, the information processing method may further
include performing a process to solve a specific task using the
model learned in the learning, wherein in the acquiring, the third
text information may be acquired from a storage apparatus in which
the third text information representing one or more utterance
sentences is stored as a test data set, in the identifying, one or
more named entities included in the acquired third text information
may be identified, in the replacing, each of the one or more
identified named entities included in the third text information
may be replaced with an abstract expression abstracted based on the
predetermined rule thereby generating fourth text information from
the third text information, and in the performing the process, the
process to solve the specific task may be performed using, as a
test data set, the fourth text information generated in the
replacing.
[0051] According to the present aspect, it is possible to use the
test data set obtained by replacing named entities included in
utterance sentences with abstract expressions in which variations
of named entities are eliminated based on the predetermined rule.
This makes it possible for the model of the dialogue system
generated by the learning to perform a process to accurately
estimate a dialogue action or a dialogue state of an utterance, and
accurately estimate or output a response content that is to be
performed by the dialogue system in response to an utterance of a
user.
[0052] For example, the information processing method may further
include performing preprocessing in which a bigram process is
performed as the prepressing to divide the second text information
and the fourth text information generated in the replacing into a
plurality of segments each including two adjacent expressions,
wherein in the learning, the model may be learned using, as
training data, the second text information subjected to the bigram
process in the preprocessing, and in the performing the process, a
further process may be performed to solve the specific task using,
as a test data set, the fourth text information subjected to the
bigram process in the preprocessing.
[0053] For example, the information processing method may further
include performing preprocessing in which a standardization process
is performed as the prepressing to standardize a format of the
expressions included in the second text information and the fourth
text information generated in the replacing, wherein in the
learning, the model may be learned using, as training data, the
second text information subjected to the standardization process in
the preprocessing, and in the performing the process, a further
process may be performed to solve the specific task using, as a
test data set, the fourth text information subjected to the
standardization process in the preprocessing.
[0054] For example, the information processing method may further
include performing outputting such that one or more abstract
expressions, included in a result of the performed process to solve
the specific task and replaced based on the predetermined rule, are
replaced with the original named entities and output.
[0055] In an aspect, the present disclosure provides an information
processing apparatus including an acquirer that acquires first text
information from a storage apparatus in which first text
information representing one or more utterance sentences is stored
as a learning data set, an identifier that identifies one or more
named entities included in the acquired first text information, a
replacement processor that replaces each of the one or more named
entities identified by the identifier with an abstract expression
abstracted based on a predetermined rule thereby generating second
text information from the first text information, and a learning
processor that learns a model of a dialogue system using, as
training data, the second text information generated by the
replacement processor.
[0056] In an aspect, the present disclosure provides a program that
causes a computer to execute the information processing method
according to the above-described aspect.
[0057] It should be noted that general or specific embodiments may
be implemented as a system, a method, an integrated circuit, a
computer program, a computer-readable storage medium such as
CD-ROM, or any selective combination thereof.
[0058] Embodiments of the present disclosure are described below
with reference to drawings. Note that each embodiment described
below is for illustrating a specific example of the present
disclosure. That is, in the following embodiments of the present
disclosure, values, shapes, constituent elements, steps, the order
of steps, and the like are described by way of example but not
limitation. Among constituent elements described in the following
embodiments, those constituent elements that are not described in
independent claims indicating highest-level concepts of the present
disclosure are optional. Also note that various combinations of
part or all of embodiments are possible.
EMBODIMENTS
[0059] Embodiments of an information processing method and related
techniques are described below with reference to drawings.
Information Processing Apparatus 10
[0060] FIG. 1 is a block diagram illustrating an example of a
configuration of the information processing apparatus 10 according
to an embodiment. FIG. 2 is a block diagram illustrating an example
of a detailed configuration of a data processor 2 shown in FIG.
1.
[0061] The information processing apparatus 10 includes, as shown
in FIG. 1, storage apparatuses 1, 4, and 5, a learning processor 3,
a prediction processor 6, and an output processor 7, and is used in
a dialogue system. This dialogue system may be a task-oriented
dialogue system.
Storage Apparatus 1, 4, and 5
[0062] The storage apparatus 1 stores a learning data set used in a
learning process by the learning processor 3. In the present
embodiment, the storage apparatus 1 stores first text information
representing one or more utterance sentences as the learning data
set used by the learning processor 3 when the learning processor 3
performs the learning process. The learning data set includes a
plurality of pieces of dialogue data each including one or more
utterance sentences.
[0063] The storage apparatus 4 stores a model to be learned by the
learning processor 3 using the learning data set. In the present
embodiment, the storage apparatus 4 stores a dialogue system model
to be learned by the learning processor 3.
[0064] The storage apparatus 5 stores a test data set used in a
prediction process by the prediction processor 6. In the present
embodiment, the storage apparatus 4 stores third text information
representing one or more utterance sentences as the test data set,
which is used by the prediction processor 6 when the prediction
processor 6 performs the prediction process. The test data set
includes a plurality of pieces of dialogue data each including one
or more utterance sentences.
[0065] The storage apparatuses 1, 4, and 5 each are realized using
a rewritable nonvolatile memory such as a hard disk drive, a solid
state drive, or the like.
Data Processor 2
[0066] The data processor 2 processes the learning data set stored
in the storage apparatus 1. In the present embodiment, the data
processor 2 generates second text information from the first text
information by replacing one or more named entities included in the
first text information with abstract expressions abstracted based
on a predetermined rule. Note that the named entity is a generic
expression of proper nouns such as person names, location names,
and the like, date expressions, time expressions, and the like.
More specifically, examples of named entities are organization
names, person names, location names, date expressions, time
expressions, money amount expressions, rate expressions, artifact
names, and the like. Note that in the present embodiment, the
abstract expressions are generalized concepts of named entities
such that different entities of abstract expressions belong to the
same concept are distinguish from each other in one dialogue, while
they are regarded identical to each other between different
dialogues.
[0067] The data processor 2 processes the test data set stored in
the storage apparatus 5. In the present embodiment, the data
processor 2 generates fourth text information from the third text
information by replacing one or more named entities included in the
third text information with abstract expressions abstracted based
on a predetermined rule.
[0068] The data processor 2 may perform a standardization process
on the generated second text information and fourth text
information to standardize the formats of expressions included
therein. The standardization process is, for example, a process of
converting a plural form of a word to a singular form, a process of
converting a past form of a word to a present form, or the like.
The data processor 2 may preform a bigram process on the generated
second text information and fourth text information to divide the
text information in units of adjacent two expressions.
[0069] A detailed configuration and related matters will be
described later.
Learning Processor 3
[0070] The learning processor 3 is realized by a computer or the
like and performs a learning process of a model of the dialogue
system based on a predetermined algorithm used in machine learning
such as deep learning or the like. The learning processor 3
performs the learning process using the learning data set processed
by the data processor 2 and stored in the storage apparatus 1. The
learning processor 3 stores the model subjected to the learning
process in the storage apparatus 4.
[0071] In the present embodiment, the learning processor 3 learns
the model of the dialogue system using, as training data, the
second text information generated by the data processor 2. The
model of the dialogue system may be a neural network model to be
subjected to machine learning and may include, for example, a
recurrent neural network.
[0072] In a case where the second text information is subjected to
the bigram process in the data processor 2, the learning processor
3 may learn the model of the dialogue system using, as training
data, the second text information subjected to the bigram process.
In a case where the second text information is subjected to the
standardization process in the data processor 2, the learning
processor 3 may learn the model of the dialogue system using, as
training data, the second text information subjected to the
standardization process.
Prediction Processor 6
[0073] The prediction processor 6 is realized by a computer or the
like. The prediction processor 6 makes the learned model stored in
the storage apparatus 4 perform a process to solve a specific task
using a test data set. More specifically, the prediction processor
6 performs this process using the test data set processed by the
data processor 2 and stored in the storage apparatus 5. The
prediction processor 6 outputs a processing result to the output
processor 7. The specific task is, for example, a restaurant
reservation task, a hotel reservation task, a burger order task, or
a rent apartment search task. However, the specific task is not
limited to these examples. The process of solving the specific task
is estimating a dialogue action or a dialogue state of an utterance
made by a user to solve a specific task, estimating and outputting
a content of a response to be made by the system in response to the
utterance made by the user, or the like. However, the process of
solving the specific task is not limited to these examples.
[0074] In the present embodiment, the prediction processor 6
performs the process to solve the specific task using the model
subjected to the learning in the learning processor 3. More
specifically, the prediction processor 6 performs the process to
solve the specific task using, as the test data, the fourth text
information generated by the data processor 2. The prediction
processor 6 then outputs a processing result to the output
processor 7.
[0075] In a case where the fourth text information is subjected to
the bigram process in the data processor 2, the prediction
processor 6 may perform the process to solve the specific task
using, as the test data, the fourth text information subjected to
the bigram process. In a case where the fourth text information is
subjected to the standardization process in the data processor 2,
the prediction processor 6 may perform the process to solve the
specific task using, as the test data, the fourth text information
subjected to the standardization process.
Output Processor 7
[0076] The output processor 7 performs an inverse conversion on the
output result acquired from the prediction processor 6 to convert
an abstract expression included in the output result to an original
named entity and outputs a conversion result. In the present
embodiment, the output processor 7 performs the inverse conversion
such that one or more abstract expressions, which have been
obtained via the replacement based on the predetermined rule and
which are included in the result of the process to solve the
specific task, are replaced with the original named entities, and
the output processor 7 outputs the result.
[0077] Details of the configuration of the data processor 2 are
described below.
Data Processor 2
[0078] The data processor 2 includes, for example, an acquirer 11,
a knowledge base 12, an identifier 13, and a replacement processor
14, as shown in FIG. 2. The data processor 2 may further include a
preprocessor (not shown in the figure) at a stage following the
replacement processor 14.
Acquirer 11
[0079] The acquirer 11 acquires the first text information from the
storage apparatus 1 in which the first text information
representing one or more utterance sentences is stored as the
learning data set. In the present embodiment, the acquirer 11
acquires the learning data set as the first text information from
the storage apparatus 1 and outputs the acquired learning data set
to the identifier 13. The acquirer 11 may acquire, as the first
text information, one dialogue at a time from the learning data
set, that is, text information representing one dialogue or may
acquire one utterance at a time from one or more utterances
included in the one dialogue, that is, text information
representing one utterance.
[0080] The acquirer 11 also acquires the third text information
from the storage apparatus 5 in which the third text information
representing one or more utterance sentences is stored as the test
data set. In the present embodiment, the acquirer 11 acquires the
test data set as the third text information from the storage
apparatus 5 and outputs the acquired test data set to the
identifier 13. The acquirer 11 may acquire, as the third text
information, one dialogue at a time from the test data set, that
is, text information representing one dialogue, or may acquire one
utterance at a time from one or more utterances included in the one
dialogue, that is, text information representing one utterance.
[0081] The acquirer 11 includes a processor such as a CPU, an ASIC,
an FPGA or the like and is realized by the processor such as the
CPU or the like by executing a computer-readable program stored in
the computer.
Knowledge Base 12
[0082] The knowledge base 12 is a database in which knowledge about
facts, common senses, and/or experiences is described in a form
interpretable by a computer. The knowledge base 12 allows it to
acquire knowledge by performing searching. Note that the knowledge
base 12 may exist outside the data processor 2. That is, the data
processor 2 does not necessarily need to include knowledge base
12.
[0083] FIG. 3 is a diagram illustrating an example of knowledge
stored in the knowledge base 12 shown in FIG. 2. In the example
shown in FIG. 3, information about restaurants is stored in the
knowledge base 12 according to the present embodiment. More
specifically, in the example shown in FIG. 3, two restaurant names
"the_sparkling_service" and "gentle_deep_water" are shown as
restaurant candidates. That is, in FIG. 3, facility names
representing target candidates, which may solve the specific task,
are indicated as examples of candidates. In FIG. 3, contents
(values) of restaurants and their attributes (slots) are also
indicated, that is, "pari" and "r_location", "cheap" and "r_price",
"40" and "r_rating", "british" and "r_cuisine", "business" and
"r_atmosphere", and "two" and "r_number" are shown. That is, in
FIG. 3, contents (values) of restaurants and their attributes
(slots), which are generalized concepts of the contents, are shown.
In other words, abstract expressions are expressions obtained by
abstracting one or more named entities according to the
predetermined rule.
Identifier 13
[0084] The identifier 13 identifies one or more named entities
included in the first text information acquired by the acquirer 11.
In the present embodiment, the identifier 13 refers to the
knowledge base 12 to identify one or more named entities that are
included in the first text information acquired by the acquirer 11
and that are to be subjected to the replacement process.
[0085] The identifier 13 also identifies one or more named entities
included in the third text information acquired by the acquirer 11.
In the present embodiment, the identifier 13 refers to the
knowledge base 12 to identify one or more named entities that are
included in the third text information acquired by the acquirer 11
and that are to be subjected to the replacement process.
Replacement Processor 14
[0086] The replacement processor 14 generates second text
information from the first text information by replacing one or
more named entities identified by the identifier 13 with abstract
expressions abstracted based on the predetermined rule.
[0087] In the present embodiment, the replacement processor 14
includes, as shown in FIG. 2, a context dictionary 15, a registerer
16, and a replacement executer 17. Note that the replacement
processor 14 may not include the context dictionary 15 and the
registerer 16.
Context Dictionary 15
[0088] The context dictionary 15 stores pairs of expressions
between which the replacement is performed depending on the
context. Using the context dictionary 15, it is possible to find
knowledge. That is, in the context dictionary 15, named entities
and corresponding abstract expressions to which the named entities
are replaced are stored.
[0089] FIG. 4 illustrates by way of example expression pairs stored
in the context dictionary 15 shown in FIG. 2. In the example shown
in FIG. 4, one or more named entities included in the first text
information or the third text information and their abstract
expressions paired with the respective named entities are stored in
the context dictionary 15 according to the present embodiment. More
specifically, in the example shown in FIG. 4, "french", "british",
"the_sparkling_service", and "gentle_deep_water" are stored in a
replacement source field. That is, as shown in FIG. 4, words or
strings of words existing in the content field (value field) or the
candidate field of the knowledge base 12 are registered in the
replacement source field of the context dictionary 15.
[0090] Furthermore, in the example shown in FIG. 4, "cuisine_1",
"cuisine_2", "cand_1", and "cand_2" are stored in the replacement
destination field. That is, as shown in FIG. 4, words or strings of
words which exist in an attributed field (a slot field) of the
knowledge base 12 and each of which has a numeral suffix added
thereto depending on a context are registered in the replacement
destination field of the context dictionary 15. The numeral values
at the suffix have serial values assigned in the order in which the
expressions occur in the value field. In other words, each numeral
value has a value assigned in the order in which the one or more
named entities occur in the one or more utterance sentences
represented in the first text information (or the third text
information).
[0091] In a case where expressions in the replacement source field
are words or strings of words existing in the candidate field,
numeral values corresponding to ranks of evaluation values on the
service provided by the facilities represented by the expression
are added as the numeral values depending on the context. The
evaluation values are values indicating evaluations about services,
such as cuisines, feelings, prices of cuisines, provided by the
facilities. The evaluation values may be determined based on
reputations, the number of comments, ranking, the number of
accesses, on the Internet or the like, or evaluation values may be
determined such that rows including "r_rating" are extracted from
the knowledge base 12 and the evaluation value may be determined
depending on the number of rows. In the example shown in FIG. 3,
the content (value) of the attribute (slot) of "r_rating" is "40"
for "the_sparkling_service", while the content (value) of the
attribute (slot) of "r_rating" is "32" for "gentle_deep_water".
Thus, "the_sparkling_service" has a higher evaluation value than
"gentle_deep_water", and thus they are respectively assigned
numbers 1 and 2. That is, "the_sparkling_service" and
"gentle_deep_water" in the replacement source field may be assigned
"cand_1" and "cand_2" in the replacement destination field.
[0092] Note that the context dictionary 15 is not limited to the
example shown in FIG. 4. For example, the numbering may not be
performed, or expressions of words or strings of words existing in
the candidate field shown in FIG. 3 may not included in the
replacement source field. That is, only expressions existing in the
context field (value field) of the knowledge base 12 may be
registered in the replacement source field. In this case,
expressions existing in the attributed field (slot field) of the
knowledge base 12 may be registered in the replacement destination
field, or expressions existing in the attributed field (slot field)
of the knowledge base 12 are added with numeral values depending on
the context, and results may be registered in the replacement
destination field.
Registerer 16
[0093] The registerer 16 includes a processor such as a CPU, an
ASIC, an FPGA or the like and is realized by the processor such as
the CPU or the like by executing a computer-readable program stored
in the computer.
[0094] In the present embodiment, the registerer 16 checks whether
the named entities, which are identified by the identifier 13 and
which are to be subjected to the replacement process, are
registered in the replacement source field and corresponding
expressions are registered in the replacement destination field of
the context dictionary 15. In other words, the registerer 16 checks
whether each of one or more named entities included in the first
text information (or third text information) is registered in the
context dictionary 15.
[0095] In a case where the registerer 16 determines that the one or
more named entities included in the first text information (or the
third text information) include a named entity that is not
registered in the context dictionary 15, the registerer 16
registers these unregistered named entities together with
corresponding expressions depending on the context in the context
dictionary 15.
Replacement Executer 17
[0096] The replacement executer 17 includes a processor such as a
CPU, an ASIC, an FPGA or the like and is realized by the processor
such as the CPU or the like by executing a computer-readable
program stored in the computer.
[0097] In the present embodiment, the replacement executer 17
performs a replacement process such that each of the one or more
named entities identified by the identifier 13 is replaced
according to the context dictionary 15. More specifically, the
replacement executer 17 replaces each of the one or more named
entities identified by the identifier 13 with an abstract
expression abstracted based on the predetermined rule, and adds a
numeral value depending on the context in the one or more utterance
sentences represented in the first text information thereby
generating second text information from the first text information.
Note that the numeral values depending on the context are assigned
in the order in which the one or more named entities occur in the
one or more utterance sentences represented in the first text
information. Furthermore, the replacement executer 17 performs a
replacement process such that, of the one or more named entities
identified by the identifier 13, facility names indicating target
candidates that may give a solution of the specific task are
replaced with abstract expressions abstracted based on the
predetermined rule, and numeral values depending on evaluation
values about the service provided by facilities indicated by the
facility names are added as the numeral values depending on the
context. The replacement executer 17 outputs the generated second
text information to the storage apparatus 1.
[0098] In some cases, only expressions existing in the context
field (value field) of the knowledge base 12 are registered in the
replacement source field of the context dictionary 15, while only
expressions existing in the attributed field (slot field) of the
knowledge base 12 are registered in the replacement destination
field. In this case, the replacement executer 17 may generate
second text information from the first text information by
replacing one or more named entities identified by the identifier
13 with abstract expressions abstracted based on a predetermined
rule. In some case, expressions existing in the attributed field
(slot field) of the knowledge base 12 are added with numeral values
depending on the context and results are registered in the
replacement destination field of the context dictionary 15. In this
case, the replacement executer 17 may replaces each of the one or
more named entities identified by the identifier 13 with abstract
expression according to the context dictionary 15 and further adds
a numeral value depending on the context in one or more utterance
sentences represented in the first text information thereby
generating the second text information from the first text
information.
[0099] The replacement process performed by the replacement
executer 17 is not limited to that in which one or more named
entity included in the first text information are respectively
replaced with abstract expressions. The replacement executer 17 may
perform the replacement process such that one or more named entity
included in the third text information are respectively replaced
with abstract expressions. For example, only expressions existing
in the context field (value field) of the knowledge base 12 may be
registered in the replacement source field of the context
dictionary 15, and only expressions existing in the attributed
field (slot field) of the knowledge base 12 may be registered in
the replacement destination field. In this case, similarly, the
replacement executer 17 may generate fourth text information from
the third text information by replacing one or more named entities
identified by the identifier 13 with abstract expressions
abstracted based on a predetermined rule. The replacement executer
17 outputs the generated fourth text information to the storage
apparatus 5. Note that the replacement executer 17 performs an
operation similarly also in the process of adding numeral values
depending on the context of the content of the context dictionary
15, and in the process of generating the fourth text information
from the third text information.
Preprocessor
[0100] The preprocessor may perform a standardization process on
the second text information and/or the fourth text information
generated by the replacement processor 14 to standardize formats of
the expressions included therein. The preprocessor may perform a
bigram process on the second text information and/or the fourth
text information generated by the replacement processor 14 to
divide the second text information and/or the fourth text
information into a plurality of segments each including two
adjacent expressions.
[0101] The preprocessor includes a processor such as a CPU, an
ASIC, an FPGA or the like and is realized by the processor such as
the CPU or the like by executing a computer-readable program stored
in the computer.
Operation of Information Processing Apparatus 10
[0102] Next, information processing performed by the information
processing apparatus 10 configured in the above-described manner is
described.
[0103] FIG. 5 is flow chart illustrating an outline of the
operation of the information processing apparatus 10 according to
an embodiment.
[0104] First, the information processing apparatus 10 performs an
acquisition process to acquire a learning data set (S1). More
specifically, the information processing apparatus 10 acquires
first text information from the storage apparatus 1 in which the
first text information representing one or more utterance sentences
is stored as the learning data set.
[0105] Next, the information processing apparatus 10 performs an
identification process to identify a named entity (S2). More
specifically, the information processing apparatus 10 identifies
one or more named entities included in the first text information
acquired in step S1.
[0106] Next, the information processing apparatus 10 performs a
replacement process to replace the named entity identified in step
S2 with an abstract expression (S3). More specifically, the
information processing apparatus 10 generates second text
information from the first text information by replacing each of
one or more named entities identified in step S2 with an abstract
expression abstracted based on a predetermined rule. Note that in
step S3, after replacing each of the one or more named entities
identified in step S2 with the abstract expression abstracted based
on the predetermined rule, the information processing apparatus 10
may add a numeral value depending on a context in one or more
utterance sentences represented in the first text information
thereby generating the second text information from the first text
information. Furthermore, in step S3, the information processing
apparatus 10 may perform a process such that a named entity, which
is included in the one or more named entities identified in step S2
and which is of a facility name indicating a target candidate that
may give a solution of the specific task, is replaced with an
abstract expression abstracted based on the predetermined rule, and
a numeral value corresponding to a rank of an evaluation value
about a service provided by a facility indicated by the facility
name is added as the numeral value depending on the context.
[0107] Next, the information processing apparatus 10 performs a
learning process (S4). More specifically, the information
processing apparatus 10 learns the model of the dialogue system
using, as training data, the second text information generated in
step S3.
[0108] Note that the process from step S1 to step S3 is performed
in a similar manner also for a test data set. In this case, the
information processing apparatus 10 may make the model already
subjected to the learning perform the process to solve the specific
task using, as the test data, the fourth text information generated
in the process in step S3 or the like.
[0109] Next, the information processing performed by the
information processing apparatus 10 is described in further detail
below.
[0110] FIG. 6A is a flow chart illustrating details of the
operation performed by the information processing apparatus 10
according to an embodiment. FIG. 6B is a flow chart illustrating
another embodiment of a process in step S103 shown in FIG. 6A. In
FIG. 6A, the operation is illustrated in detail for a case where
expressions existing in the candidate field shown in FIG. 3 are not
included in the replacement source field of the context dictionary
15. On the other hand, the operation shown in FIG. 6B, unlike the
operation shown in FIG. 6A, is shown for a case where expressions
existing in the candidate field shown in FIG. 3 are included in the
replacement source field of the context dictionary 15.
[0111] First, as shown in FIG. 6A, the information processing
apparatus 10 acquires a data set (S101). The data set may be a
learning data set or a test data set. In a case where the data set
is a learning data set, the information processing apparatus 10
acquires the learning data set from the storage apparatus 1. On the
other hand, in a case where the data set is a test data set, the
information processing apparatus 10 acquires the test data set from
the storage apparatus 5.
[0112] Next, the information processing apparatus 10 acquires one
dialogue from the data set acquired in step S101 (S102).
[0113] Next, the information processing apparatus 10 initializes
the context dictionary 15 (S103). More specifically, the
information processing apparatus 10 clears the context dictionary
15 by clearing all expression pairs stored in the context
dictionary 15.
[0114] Next, the information processing apparatus 10 acquires one
utterance from the one dialogue acquired in step S102 (S104).
[0115] Next, the information processing apparatus 10 checks whether
each named entity included in the one utterance acquired in step
S104 exists in the knowledge base 12 (S105). More specifically, the
information processing apparatus 10 refers to the knowledge base 12
and checks whether one or more named entities included in the one
utterance acquired in step S104 exist in the knowledge base 12.
[0116] In a case where the information processing apparatus 10
determines in step S105 that at least one of the one or more named
entities included in the one utterance acquired in step S104 exists
in the knowledge base 12 (Yes in S105), the process proceeds to
step S106. On the other hand, in a case where the information
processing apparatus 10 determines in step S105 that any one of the
one or more named entities included in the one utterance acquired
in step S104 does not exist in the knowledge base 12 (No in S105),
the process returns to step S104 in which the information
processing apparatus 10 acquires next one utterance from the one
dialogue acquired in step S102.
[0117] Next, the information processing apparatus 10 checks whether
a named entity that is included in the one or more named entities
included in the one utterance acquired in step S104 and that exists
in the knowledge base 12 is unregistered in the context dictionary
15 (S106).
[0118] In a case where the information processing apparatus 10
determines in step S106 that named entities existing in the
knowledge base 12 are all registered in the context dictionary 15
(No in S106), the process proceeds to step S108. On the other hand,
in a case where the information processing apparatus 10 determines
in step S106 that some named entity existing in the knowledge base
12 is unregistered in the context dictionary 15 (Yes in S106), the
process proceeds to step S107.
[0119] Next, in step S107, the information processing apparatus 10
registers, in the context dictionary 15, the named entity
unregistered in the context dictionary 15. More specifically, if
there is a named entity that is included in the named entities
existing in the knowledge base 12 and that is unregistered in the
context dictionary 15, then the information processing apparatus 10
registers, in the context dictionary 15, a pair of the unregistered
named entity and a context-dependent expression corresponding to
this named entity. More specifically, as shown in FIG. 4, the
information processing apparatus 10 registers the named entity in
the context dictionary 15 such that the named entity unregistered
in the context dictionary 15 is described in the replacement source
field and the context-dependent expression paired with the
unregistered named entity is described in the replacement
destination field. Note that the expression described in the
replacement destination field is an abstract expression obtained by
abstracting the named entity described in the replacement source
field by concept-generalization or the like based on the
predetermined rule.
[0120] Next, in step S108, the information processing apparatus 10
performs a replacement process such that a named entity that is
included in the one or more named entities included in the one
utterance acquired in step S104 and that exists in the knowledge
base 12 is replaced according to the context dictionary 15. More
specifically, the information processing apparatus 10 performs the
replacement process such that the named entity included in the one
utterance acquired in step S104 is replaced with an abstract
expression according to the context dictionary 15 and is
furthermore added with a numeral value depending on a context. Note
that the numeral value depending on the context is, as described
above, assigned in the order in which the one or more named
entities occur in the one or more utterance sentences.
[0121] Next, the information processing apparatus 10 checks whether
the one utterance subjected to the replacement process in step S108
is a last one included in the one piece of dialogue data acquired
in step S102 (S109).
[0122] In a case where the information processing apparatus 10
determines in step S109 that the one utterance subjected to the
replacement process in step S108 is the last one included in the
one piece of dialogue data acquired in step S102 (Yes in S109), the
process proceeds to step S110. On the other hand, in a case where
the information processing apparatus 10 determines in step S109
that the one utterance subjected to the replacement process in step
S108 is not the last one included in the one piece of dialogue data
acquired in step S102 (No in S109), the process returns to step
S104 in which the information processing apparatus 10 acquires next
one utterance from the one dialogue acquired in step S102.
[0123] Next, in step S110, the information processing apparatus 10
checks whether the one piece of dialogue data acquired in step S102
is a last one in the data set acquired in step S101.
[0124] In a case where the information processing apparatus 10
determines in step S110 that the one piece of dialogue data
acquired in step S102 is the last one in the data set acquired in
step S101 (Yes in S110), the process proceeds to step S111. On the
other hand, in a case where the information processing apparatus 10
determines in step S110 that the one piece of dialogue data
acquired in step S102 is not the last one in the data set acquired
in step S101 (No in S110), the process returns to step S102 in
which the information processing apparatus 10 acquires next one
piece of dialogue data from the data set acquired in step S101.
[0125] In step S111, the information processing apparatus 10
outputs the data set subjected to the replacement process (S111).
In a case where the data set is a learning data set, the
information processing apparatus 10 stores the learning data set
subjected to the replacement process in the storage apparatus 1. On
the other hand, in a case where the data set is a test data set,
the information processing apparatus 10 stores the test data set
subjected to the replacement process in the storage apparatus
5.
[0126] Next, referring to FIG. 6B, an operation is described for a
case where the replacement source field of the context dictionary
15 includes an expression included in the candidate field shown in
FIG. 3. The process shown in FIG. 6B is different from the process
shown in FIG. 6A in that the process in step S103 is replaced with
step S103A.
[0127] More specifically, in step S103A, first, the information
processing apparatus 10 clears the context dictionary 15 by
clearing all expression pairs stored in the context dictionary 15
(S1031).
[0128] Next, the information processing apparatus 10 checks whether
a candidate name included in the knowledge base 12 exists in the
one dialogue acquired in step S102 (S1032). More specifically, the
information processing apparatus 10 refers to the knowledge base 12
and checks whether the one dialogue acquired in step S102 includes
a named entity existing in the candidate field of the knowledge
base 12. Herein, each named entity existing in the candidate field
of the knowledge base 12 is referred to as a candidate name. The
information processing apparatus 10 is capable of checking whether
there is a named entity existing in the candidate field of the
knowledge base 12, by, for example, setting a search condition for
the one dialogue acquired in step S102 and searching the knowledge
base 12.
[0129] In a case where the information processing apparatus 10
determines in step S1032 that a candidate name included in the
knowledge base 12 exists in the one dialogue acquired in step S102
(Yes in S1032), the information processing apparatus 10 extracts
the candidate name (S1033). In a case where the information
processing apparatus 10 determines in step S1032 that a candidate
name included in the knowledge base 12 does not exist in the one
dialogue acquired in step S102 (No in S1032), the information
processing apparatus 10 proceeds to step S104 shown in FIG. 6A to
continue the process.
[0130] Next, the information processing apparatus 10 ranks the
candidate names extracted in step S1033 (S1034). For example, the
information processing apparatus 10 ranks the candidate names
according to evaluation values in terms of service provided by
facilities indicated by the respective candidate names.
[0131] Next, the information processing apparatus 10 registers, in
the context dictionary 15, the candidate names extracted in step
S1033 (S1035). More specifically, the information processing
apparatus 10 registers the candidate names extracted in step S1033
in the replacement source field of the context dictionary 15 and
registers expressions obtained by adding numeral values to "cand"
in the replacement destination field. In the registration
operation, the ranks determined in step S1034 are employed as the
numeral value.
[0132] In the manner described above, the information processing
apparatus 10 makes the registration in step S103A such that
expressions existing in the candidate field shown in FIG. 3 are
registered in the replacement source field of the context
dictionary 15.
[0133] Next, specific embodiments are described below as to the
replacement process in one dialogue in a data set. In first to
fourth embodiments, examples are described for a case where any
expression existing in the candidate field of the knowledge base 12
is not included in the replacement source field of the context
dictionary 15. In fifth to seventh embodiments, examples are
described for a case where the replacement source field of the
context dictionary 15 includes only expressions existing in the
candidate field of the knowledge base 12.
First Embodiment
[0134] In the first embodiment described below, it is assumed by
way of example that the dialogue system is a task-oriented dialogue
system configured to provide a restaurant reservation service. In
other words, this restaurant reservation dialogue system is an
example of a dialogue system in which the specific task to be
solved is a restaurant reservation task.
[0135] FIG. 7 is a diagram illustrating an example of one dialogue
which is in a state in which a replacement process according to the
first embodiment is not yet performed on the one dialogue. FIG. 8
is a diagram illustrating an example of the context dictionary 15
according to the first embodiment. FIG. 9 is a diagram illustrating
an example of one dialogue which is in a state in which the one
dialogue has been subjected to the replacement process according to
the first embodiment.
[0136] Note that the one dialogue shown in FIG. 7 may be one
dialogue included in the learning data set acquired from the
storage apparatus 1 or may be one dialogue included in the test
data set acquired from the storage apparatus 5. In FIG. 7 and FIG.
9, a symbol "U:" indicates that an utterance following this symbol
is an utterance of a user, while a symbol "S:" indicates that an
utterance following this symbol is an utterance of the dialogue
system. In the one dialogue shown in FIG. 7, a user utters about a
condition of a restaurant. The dialogue system utters a plurality
of times until the dialogue system acquires all attributes (slots).
When all attributes (slots) are obtained, the dialogue system
performs api_call to access, for example, a restaurant DB managed
by the dialogue system.
[0137] The information processing apparatus 10 refers to the
knowledge base 12 shown in FIG. 3 and checks whether a plurality of
named entities included in text information representing the one
dialogue shown in FIG. 7 exist in the knowledge base 12. The one
dialogue shown in FIG. 7 includes named entities relating to the
restaurant reservation, such as "business", "cheap", "pads",
"french", "british", and the like which exist in the knowledge base
12.
[0138] Thus, the information processing apparatus 10 registers
these named entities in the replacement source field of the context
dictionary 15 as shown in FIG. 8. The information processing
apparatus 10 makes a further registration such that abstract
expressions "atmosphere_1", "price_1", "location_1", "cuisine_1",
and "cuisine_2" corresponding to the respective named entities are
registered in the replacement destination field of the context
dictionary 15. Note that each of these abstract expressions
includes a numeral value added as a suffix assigned in the order in
which the abstract expression corresponding to the named entity
occurs in the attributed field (slot field) of the knowledge base
12.
[0139] Thus, the information processing apparatus 10 is capable of
obtaining text information representing one dialogue such as that
shown in FIG. 9 by performing the replacement process on the one
dialogue sentence shown in FIG. 7 according to the context
dictionary 15 shown in FIG. 8. As can be seen from FIG. 9, named
entities "business", "cheap", "paris", "french", and "british" are
respectively replaced with abstract expressions "atmosphere_1",
"price_1", "location_1", "cuisine_1", and "cuisine_2".
[0140] According to the present embodiment, the text information
representing a one dialogue relating to the restaurant reservation
subjected to the replacement process is allowed to be used as
training data or a test data set. That is, this provides such
training data or the test data in which different entities in
abstract expressions belonging to the same concept are
distinguished in one dialogue but are not distinguished between
different dialogues.
[0141] By learning the model of the restaurant reservation system
using the training data obtained in the above-described manner, an
improvement in the accuracy of the model of the restaurant
reservation system generated by the learning is achieved.
[0142] Use of the test data set generated in the above-described
manner makes it possible for the model of the restaurant
reservation system generated by the learning to perform a process
to accurately estimate a dialogue action or a dialogue state of an
utterance, and accurately estimate a response content to be
performed in response to an utterance of a user and output the
response.
Second Embodiment
[0143] In the second embodiment described below, it is assumed by
way of example that the dialogue system is a task-oriented dialogue
system configured to provide a hotel reservation service. In other
words, this hotel reservation system is an example of a dialogue
system in which the specific task to be solved is a hotel
reservation task. The following description of the present
embodiment focuses differences from the first embodiment.
[0144] FIG. 10 is a diagram illustrating an example of one dialogue
which is in a state in which a replacement process according to the
second embodiment is not yet performed on the one dialogue. FIG. 11
is a diagram illustrating an example of the context dictionary 15
according to the second embodiment. FIG. 12 is a diagram
illustrating an example of one dialogue which is in a state in
which the one dialogue has been subjected to the replacement
process according to the second embodiment.
[0145] The information processing apparatus 10 refers to the
knowledge base 12 and checks whether a plurality of named entities
included in text information representing the one dialogue shown in
FIG. 10 exist in the knowledge base 12. As can be seen from FIG.
10, named entities relating to the hotel reservation, such as
"vacation", "two", "london", "expensive", "cheap" and the like
which exist in the knowledge base 12.
[0146] Thus, the information processing apparatus 10 registers
these named entities in the replacement source field of the context
dictionary 15 as shown in FIG. 11. Furthermore, the information
processing apparatus 10 makes a further registration such that
abstract expressions "atmosphere_1", "number_1", "location_1",
"price_1", "price_2", corresponding to the respective named
entities are registered in the replacement destination field of the
context dictionary 15. Note that each of these abstract expressions
includes a numeral value added as a suffix assigned in the order in
which the abstract expression corresponding to the named entity
occurs in the attributed field (slot field) of the knowledge base
12.
[0147] Thus, the information processing apparatus 10 is capable of
obtaining text information representing one dialogue such as that
shown in FIG. 12 by performing the replacement process on the one
dialogue sentence shown in FIG. 10 according to the context
dictionary 15 shown in FIG. 11. As can be seen from FIG. 12, named
entities "vacation", "two", "london", "expensive", and "cheap", are
respectively replaced with abstract expressions "atmosphere_1",
"number_1", "location_1", "price_1", and "price_2".
[0148] According to the present embodiment, the text information
representing a one dialogue relating to the hotel reservation
subjected to the replacement process is allowed to be used as
training data or a test data set. That is, this provides such
training data or the test data in which different entities in
abstract expressions belonging to the same concept are
distinguished in one dialogue but are not distinguished between
different dialogues.
[0149] Thus, an improvement in the accuracy of the model of the
hotel reservation system generated by the learning is achieved.
Furthermore, it becomes possible for the model of the hotel
reservation system generated by the learning to perform a process
to accurately estimate a dialogue action or a dialogue state of an
utterance and accurately estimate or output a response content to
an utterance of a user.
Third Embodiment
[0150] In the second embodiment described below, it is assumed by
way of example that the dialogue system is a task-oriented dialogue
system configured to provide a burger order service. In other
words, the burger order system is an example of a dialogue system
in which the specific task to be solved is a burger order task. The
following description of the present embodiment focuses differences
from the previous embodiments described above.
[0151] FIG. 13 is a diagram illustrating an example of one dialogue
which is in a state in which a replacement process according to the
third embodiment is not yet performed on the one dialogue. FIG. 14
is a diagram illustrating an example of the context dictionary 15
according to the third embodiment. FIG. 15 is a diagram
illustrating an example of one dialogue which is in a state in
which the one dialogue has been subjected to the replacement
process according to the third embodiment.
[0152] The information processing apparatus 10 refers to the
knowledge base 12 and checks whether a plurality of named entities
included in text information representing the one dialogue shown in
FIG. 13 exist in the knowledge base 12. The one dialogue shown in
FIG. 13 includes named entities relating to the burger order, such
as "chicken", "onion", "mustard", "wheat", "cheese", and the like
which exist in the knowledge base 12.
[0153] Thus, the information processing apparatus 10 registers
these named entities in the replacement source field of the context
dictionary 15 as shown in FIG. 14. The information processing
apparatus 10 makes a further registration such that abstract
expressions "main_1", "vagetable_1", "sauce_1", "bun_1", "bun_2"
corresponding to the respective named entities are registered in
the replacement destination field of the context dictionary 15.
Note that each of these abstract expressions includes a numeral
value added as a suffix assigned in the order in which the abstract
expression corresponding to a named entity occurs in the attributed
field (slot field) of the knowledge base 12.
[0154] Thus, the information processing apparatus 10 is capable of
obtaining text information representing one dialogue such as that
shown in FIG. 15 by performing the replacement process on the one
dialogue sentence shown in FIG. 13 according to the context
dictionary 15 shown in FIG. 14. As can be seen from FIG. 15, named
entities "chicken", "onion", "mustard", "wheat", and "cheese" are
respectively replaced with abstract expressions "main_1",
"vagetable_1", "sauce_1", "bun_1", and "bun_2".
[0155] According to the present embodiment, the text information
representing a one dialogue relating to the burger order subjected
to the replacement process is allowed to be used as training data
or a test data set. That is, this provides such training data or
the test data in which different entities in abstract expressions
belonging to the same concept are distinguished in one dialogue but
are not distinguished between different dialogues.
[0156] Thus, an improvement in the accuracy of the model of the
burger order system generated by the learning is achieved.
Furthermore, it becomes possible for the model of the burger order
system generated by the learning to perform a process to accurately
estimate a dialogue action or a dialogue state of an utterance and
accurately estimate or output a response content in response to an
utterance of a user.
Fourth Embodiment
[0157] In the second embodiment described below, it is assumed by
way of example that the dialogue system is a task-oriented dialogue
system configured to provide a rent apartment search service. In
other words, the rent apartment search system is an example of a
dialogue system in which the specific task to be solved is a rent
apartment search task. The following description of the present
embodiment focuses differences from the previous embodiments
described above.
[0158] FIG. 16 is a diagram illustrating an example of one dialogue
which is in a state in which a replacement process according to the
fourth embodiment is not yet performed on the one dialogue. FIG. 17
is a diagram illustrating an example of the context dictionary 15
according to the fourth embodiment. FIG. 18 is a diagram
illustrating an example of one dialogue which is in a state in
which the one dialogue has been subjected to the replacement
process according to the fourth embodiment.
[0159] The information processing apparatus 10 refers to the
knowledge base 12 and checks whether a plurality of named entities
included in text information representing the one dialogue shown in
FIG. 16 exist in the knowledge base 12. The one dialogue shown in
FIG. 16 includes named entities relating to the rent apartment
search, "three", "new", "cheap", "setagaya", "shinjuku" and the
like which exist in the knowledge base 12.
[0160] Thus, the information processing apparatus 10 registers
these named entities in the replacement source field of the context
dictionary 15 as shown in FIG. 17. The information processing
apparatus 10 makes a further registration such that abstract
expressions "number_1", "age_1", "price_1", "location_1",
"location_2" corresponding to the respective named entities are
registered in the replacement destination field of the context
dictionary 15. Note that each of these abstract expressions
includes a numeral value added as a suffix assigned in the order in
which the abstract expression corresponding to named entity occurs
in the attributed field (slot field) of the knowledge base 12.
[0161] Thus, the information processing apparatus 10 is capable of
obtaining text information representing one dialogue such as that
shown in FIG. 18 by performing the replacement process on the one
dialogue sentence shown in FIG. 16 according to the context
dictionary 15 shown in FIG. 17. As can be seen from FIG. 18, named
entities "three", "new", "cheap", "setagaya", and "shinjuku" are
respectively replaced with abstract expressions "number_1",
"age_1", "price_1", "location_1", and "location_2".
[0162] According to the present embodiment, the text information
representing a one dialogue relating to the rent apartment search
subjected to the replacement process is allowed to be used as
training data or a test data set. That is, this provides such
training data or the test data in which different entities in
abstract expressions belonging to the same concept are
distinguished in one dialogue but are not distinguished between
different dialogues.
[0163] Thus, an improvement in the accuracy of the model of the
rent apartment search system generated by the learning is achieved.
Furthermore, it becomes possible for the model of the rent
apartment search system generated by the learning to perform a
process to accurately estimate a dialogue action or a dialogue
state of an utterance and accurately estimate or output a response
content in response to an utterance of a user.
Fifth Embodiment
[0164] In the fifth embodiment described below, it is assumed by
way of example that the dialogue system is a task-oriented dialogue
system configured to provide a restaurant reservation service.
[0165] FIG. 19 is a diagram illustrating an example of one dialogue
which is in a state in which a replacement process according to the
fifth embodiment is not yet performed on the one dialogue. FIG. 20
is a diagram illustrating a part of a knowledge base 12 according
to the fifth embodiment. FIG. 21 is a diagram illustrating an
example of the context dictionary 15 according to the fifth
embodiment. FIG. 22 is a diagram illustrating an example of one
dialogue which is in a state in which the one dialogue has been
subjected to the replacement process according to the fifth
embodiment.
[0166] Note that the one dialogue shown in FIG. 19 may be one
dialogue included in the learning data set acquired from the
storage apparatus 1 or may be one dialogue included in the test
data set acquired from the storage apparatus 5. In FIG. 19 and FIG.
22, a symbol "U:" indicates that an utterance following this symbol
is an utterance of a user, while a symbol "S:" indicates that an
utterance following this symbol is an utterance of the dialogue
system. In the one dialogue shown in FIG. 19, the dialogue system
gives utterances to propose to a user restaurant candidates in the
order of evaluation from the highest to the lowest until the user
accepts one of the proposed restaurant candidates.
[0167] The information processing apparatus 10 refers to the
knowledge base 12 shown in FIG. 20 and checks whether candidate
names described in the knowledge base 12 exist in the text
information representing the one dialogue shown in FIG. 19. Note
that the candidate names described in the knowledge base 12 are
facility names indicating target candidates that may solve the
specific task, which is the restaurant reservation in this case,
and the facility names are named entities indicating restaurant
names. The candidate names in the knowledge base 12 correspond to
named entities existing in the candidate field of the knowledge
base 12 shown in FIG. 3. The one dialogue shown in FIG. 19 includes
named entities representing restaurant names, such as
"the_sparkling_service", "gentle_deep_water", and the like which
exist in the knowledge base 12 shown in FIG. 20.
[0168] Thus, the information processing apparatus 10 registers
these named entities representing these restaurant names in the
replacement source field of the context dictionary 15 as shown in
FIG. 21. The information processing apparatus 10 makes a further
registration such that abstract expressions "cand_1" and "cand_2"
corresponding to the respective named entities representing the
restaurant names are registered in the replacement destination
field of the context dictionary 15. Note that each of these
abstract expressions includes a numeral value added as a suffix
assigned in the descending order of the evaluation values for the
named entities indicating the restaurant names.
[0169] In the knowledge base 12 shown in FIG. 20, "40" is stored as
a value (that is, a content) of the attribute (slot) of "r_rating"
corresponding to "the_sparkling_service". On the other hand, "32"
is stored as a value (that is, a content) of the attribute (slot)
of "r_rating" corresponding to the "gentle_deep_water". Thus the
evaluation value for "the_sparkling_service" is hither than the
evaluation value for "gentle_deep_water". Thus, in the context
dictionary 15, "cand_1" is registered as the abstract expression
corresponding to "the_sparkling_service", while "cand_2" is
registered as the abstract expression corresponding to
"gentle_deep_water".
[0170] Thus, the information processing apparatus 10 is capable of
obtaining text information representing one dialogue such as that
shown in FIG. 22 by performing the replacement process on the one
dialogue sentence shown in FIG. 19 according to the context
dictionary 15 shown in FIG. 21. As can be seen from FIG. 22, named
entities "the_sparkling_service", and "gentle_deep_water" are
respectively replaced with abstract expressions "cand_1", and
"cand_2".
[0171] According to the present embodiment, the text information
representing a one dialogue relating to the restaurant reservation
subjected to the replacement process is allowed to be used as
training data or a test data set. That is, this provides such
training data or the test data in which different entities in
abstract expressions belonging to the same concept are
distinguished in one dialogue but are not distinguished between
different dialogues.
[0172] By learning the model of the restaurant reservation system
using the training data obtained in the above-described manner, an
improvement in the accuracy of the model of the restaurant
reservation system generated by the learning is achieved.
[0173] Use of the test data set generated in the above-described
manner makes it possible for the model of the restaurant
reservation system generated by the learning to perform a process
to accurately estimate a dialogue action or a dialogue state of an
utterance and accurately estimate or output a response content to
an utterance of a user.
Sixth Embodiment
[0174] In the sixth embodiment described below, it is assumed by
way of example that the dialogue system is a task-oriented dialogue
system configured to provide a hotel reservation service. The
following description of the present embodiment focuses differences
from the fifth embodiment.
[0175] FIG. 23 is a diagram illustrating an example of one dialogue
which is in a state in which a replacement process according to the
sixth embodiment is not yet performed on the one dialogue. FIG. 24
is a diagram illustrating a part of a knowledge base 12 according
to the sixth embodiment. FIG. 25 is a diagram illustrating an
example of the context dictionary 15 according to the sixth
embodiment. FIG. 26 is a diagram illustrating an example of one
dialogue which is in a state in which the one dialogue has been
subjected to the replacement process according to the sixth
embodiment.
[0176] The information processing apparatus 10 refers to the
knowledge base 12 shown in FIG. 24 and checks whether candidate
names described in the knowledge base 12 exist in the text
information representing the one dialogue shown in FIG. 23. Note
that the candidate names described in the knowledge base 12 are
facility names indicating target candidates that may solve the
specific task, which is the hotel reservation in this case, and the
facility names are named entities indicating hotel names. The one
dialogue shown in FIG. 23 includes named entities representing
hotel names, such as "green_dog_garden", "deep_wolf_hill", and the
like which exist in the knowledge base 12 shown n FIG. 24.
[0177] Thus, the information processing apparatus 10 registers
these named entities representing these hotel names in the
replacement source field of the context dictionary 15 as shown in
FIG. 25. The information processing apparatus 10 makes a further
registration such that abstract expressions "cand_1" and "cand_2"
corresponding to the respective named entities representing the
hotel names are registered in the replacement destination field of
the context dictionary 15. Note that each of these abstract
expressions includes a numeral value added as a suffix assigned in
the descending order of the evaluation value for the named entity
indicating the hotel name.
[0178] In the knowledge base 12 shown in FIG. 24, "55" is stored as
a value of "r_rating" corresponding to "green_dog_garden", while
"34" is stored as a value of "r_rating" corresponding to
"deep_wolf_hill". Thus, the evaluation value for "green_dog_garden"
is higher than the evaluation value for "deep_wolf_hill". Thus, in
the context dictionary 15, "cand_1" is registered as the abstract
expression corresponding to "green_dog_garden", while "cand_2" is
registered as the abstract expression corresponding to
"deep_wolf_hill".
[0179] Thus, the information processing apparatus 10 is capable of
obtaining text information representing one dialogue such as that
shown in FIG. 26 by performing the replacement process on the one
dialogue sentence shown in FIG. 23 according to the context
dictionary 15 shown in FIG. 25. As can be seen from FIG. 26, named
entities "green_dog_garden", and "deep_wolf_hill" are respectively
replaced with abstract expressions "cand_1", and "cand_2".
[0180] According to the present embodiment, the text information
representing a one dialogue relating to the hotel reservation
subjected to the replacement process is allowed to be used as
training data or a test data set. That is, this provides such
training data or the test data in which different entities in
abstract expressions belonging to the same concept are
distinguished in one dialogue but are not distinguished between
different dialogues.
[0181] Thus, an improvement in the accuracy of the model of the
hotel reservation system generated by the learning is achieved.
Furthermore, it becomes possible for the model of the hotel
reservation system generated via the learning to accurately
estimate a dialogue action or a dialogue state of an utterance and
estimate and output a content in response to the utterance of a
user.
Seventh Embodiment
[0182] In the seventh embodiment described below, it is assumed by
way of example that the dialogue system is a task-oriented dialogue
system configured to provide a rent apartment search service. The
following description of the present embodiment focuses differences
from the fifth embodiment and other embodiments.
[0183] FIG. 27 is a diagram illustrating an example of one dialogue
which is in a state in which a replacement process according to the
seventh embodiment is not yet performed on the one dialogue. FIG.
28 is a diagram illustrating a part of a knowledge base 12
according to the seventh embodiment. FIG. 29 is a diagram
illustrating an example of the context dictionary 15 according to
the seventh embodiment. FIG. 30 is a diagram illustrating an
example of one dialogue which is in a state in which the one
dialogue has been subjected to the replacement process according to
the seventh embodiment.
[0184] The information processing apparatus 10 refers to the
knowledge base 12 shown in FIG. 28 and checks whether candidate
names described in the knowledge base 12 exist in the text
information representing the one dialogue shown in FIG. 27. Note
that the candidate names described in the knowledge base 12 are
facility names indicating target candidates that may solve the
specific task, which is the rent apartment search in this case, and
the facility names are named entities indicating names of buildings
including rent apartments. The one dialogue shown in FIG. 27
includes named entities representing name of a building including a
rent apartments, such as "grand_union_forest",
"water_colorful_tree", and the like which exist in the knowledge
base 12 shown in FIG. 28.
[0185] Thus, the information processing apparatus 10 registers
these named entities representing these building names in the
replacement source field of the context dictionary 15 as shown in
FIG. 29. The information processing apparatus 10 makes a further
registration such that abstract expressions "cand_1" and "cand_2"
corresponding to the respective named entities representing the
building names are registered in the replacement destination field
of the context dictionary 15. Note that each of these abstract
expressions includes a numeral value added as a suffix assigned in
the descending order of the evaluation value for the named entity
indicating the building.
[0186] In the knowledge base 12 shown in FIG. 28, "0.8" is stored
as a value of "r_distance" corresponding to "grand_union_forest",
while "1.2" is stored as a value of "r_distance" corresponding to
"water_colorful_tree". Thus, the building including the rent
apartment indicated by "grand_union_forest" is closer than the
building including the rent apartment indicated by
"water_colorful_tree", and thus the former has a higher evaluation
value. Thus, in the context dictionary 15, "cand_1" is registered
as the abstract expression corresponding to "grand_union_forest",
while "cand_2" is registered as the abstract expression
corresponding to "water_colorful_tree".
[0187] Thus, the information processing apparatus 10 is capable of
obtaining text information representing one dialogue such as that
shown in FIG. 30 by performing the replacement process on the one
dialogue sentence shown in FIG. 27 according to the context
dictionary 15 shown in FIG. 25. As can be seen from FIG. 30, named
entities "grand_union_forest" and "water_colorful_tree" are
respectively replaced with abstract expressions "cand_1" and
"cand_2".
[0188] According to the present embodiment, the text information
representing a one dialogue relating to the rent apartment search
subjected to the replacement process is allowed to be used as
training data or a test data set. That is, this provides such
training data or the test data in which different entities in
abstract expressions belonging to the same concept are
distinguished in one dialogue but are not distinguished between
different dialogues.
[0189] Thus, an improvement in the accuracy of the model of the
rent apartment search system generated by the learning is achieved.
Furthermore, it becomes possible for the model of the rent
apartment search system generated by the learning to perform a
process to accurately estimate a dialogue action or a dialogue
state of an utterance and accurately estimate or output a response
content in response to an utterance of a user.
EFFECTS AND RELATED ISSUES
[0190] As described above, according to embodiments, it is possible
to learn the model of the dialogue system using training data
obtained by replacing named entities included in utterance
sentences with abstract expressions in which variations of named
entities are eliminated based on the predetermined rule. Thus, in
the information processing method and the information processing
apparatus, an improvement in the accuracy of the model of the
dialogue system generated by the learning is achieved.
[0191] Note that in a case where an utterance sentence includes
named entities which will fall within the same concept when they
are abstracted, different entities belonging to the same concept
are regarded as identical to each other between different
dialogues, but they cannot be distinguished from each other in one
dialogue. This may cause a reduction in accuracy of the model of
the dialogue system generated via learning.
[0192] To handle the above situation, numeral values depending on
the context may be added to the abstract expressions obtained by
the replacement based on the predetermined rule. That is, according
to the present embodiment, in addition to replacing named entities
included in utterance sentence with abstract expressions, numeral
values depending on a context may be added to the abstract
expressions. This makes it possible to provide training data in
which different entities in abstract expressions belonging to the
same concept are distinguished in one dialogue but they are
regarded as identical to each other between different dialogues,
and thus a further improvement in the accuracy of the model of the
dialogue system generated by the learning is achieved.
[0193] According to the present embodiment, training data obtained
by replacing named entities included in an utterance sentence with
abstract expressions may be further subjected to a bigram process
or a standardization process as preprocessing, and then a dialogue
system model may be learned. This allows it to achieve a further
improvement in the accuracy of the model of the dialogue system
generated by the learning.
[0194] According to the present embodiment, similarly, a test data
set may be used which is obtained by replacing named entities
included in an utterance sentence with abstract expressions in
which variations of named entities are eliminated. This makes it
possible for the model of the dialogue system generated by the
learning to perform a process to accurately estimate a dialogue
action or a dialogue state of an utterance and accurately estimate
or output a response content that is to be performed by the
dialogue system in response to an utterance of a user.
[0195] According to the present embodiment, test data obtained by
replacing named entities included in an utterance sentence with
abstract expressions may be further subjected to a bigram process
or a standardization process as preprocessing, and then a result
may be processed by a dialogue system model generated via learning.
This makes it possible for the model of the dialogue system
generated by the learning to perform a process to more accurately
estimate a dialogue action or a dialogue state of an utterance and
more accurately estimate or output a response content that is to be
performed by the dialogue system in response to an utterance of a
user.
Experimental Results
[0196] A dialogue system model was obtained via learning using a
learning data set obtained by performing a replacement process
using an information processing apparatus according to an
information processing method of the present disclosure, and
effectiveness was evaluated for a case where the dialogue system
estimated a response content to be output by the dialogue system.
Experimental results were as follows.
Experimental Conditions
[0197] In an experiment, End-to-End Memory Networks disclosed in
Sukhbaatar, S., Szlam, A., Weston, J., and Fergus, R. "End-to-end
memory networks", Proceedings of NIPS, R, (2015) were used in a
dialogue system model such as a restaurant reservation system.
Learning was performed under a condition in which a learning rate,
an epoch, the number of hops, and Embedding Dim were respectively
set to 0.01, 200, 2, and 32, and the other parameters were set to
values described in Sukhbaatar, S., Szlam, A., Weston, J., and
Fergus, R. "End-to-end memory networks", Proceedings of NIPS, R,
(2015). In the knowledge base 12, cuisines, locations, price zones,
the number of people for reservation, feelings, restriction
handling, addresses, and phone numbers were stored via
learning.
[0198] Two types test data sets, that is, a data set for test #1
and a data set for test #2 each including 1000 pieces of data were
prepared. The data set for test #1 and the data set for test #2
each include an utterance of a user and a plurality of utterances
selectable by the dialogue system in response to the utterance.
When the dialogue system selected an utterance to be selected in
response to an utterance of a user according to the dialogue system
model, it was evaluated that the selection was correct.
[0199] For example, in a case where one dialogue includes six
utterances as in the example shown in FIG. 7, the data set for test
#1 and the data set for test #2 each include five utterances except
for the last 6th utterance, and include an utterance to be the 6th
utterance as a choice. In a case where the dialogue system selected
the 6th utterance in the one dialogue shown in FIG. 7, it was
evaluated that the correct selection was made by the performance of
the dialogue system. More specifically, the performance of the
dialogue system model was given by the ratio of corrected
selections made by the dialogue system model for utterances of a
user with respect to the total number (1000 in this specific
example) of pieces of data included in each of the data set for
test #1 and the data set for test #2.
[0200] Note that the data set for test #1 and the data set for test
#2 were different from each other in terms of the cuisine and the
location. Another difference was in that the data set for test #2
included unknown words. In this experiment, the learning was
performed using the data set for test #1.
Experimental Results
[0201] FIGS. 31 to 33 illustrates effects of the replacement
process according to the present disclosure. In each of FIGS. 31 to
33, "Original" indicates experimental results, in a comparative
example, obtained for a case where the model used was obtained via
the learning without performing the replacement process according
to the present disclosure.
First Experimental Result
[0202] FIG. 31 shows a first experimental result for a case where a
task similar to the task for the one dialogue shown in the first
embodiment was performed. In the task of this experiment, a user
uttered about a condition as to a restaurant, and the dialogue
system made utterances until all attributes (slots) were acquired
from the user. When all attributes (slots) were acquired, the
dialogue system performed api_call to perform a task to access, for
example, a restaurant DB managed by the dialogue system.
[0203] In FIG. 31, "+replacement" shows performances obtained
respectively using the data set for test #1 and the data set for
test #2 subjected to the replacement such that named entities were
replaced with abstract expressions and numeral values were added
thereto in accordance with the order occurrence in the dialogue. In
FIG. 31, "+replacement, bi-gram, removing candidate" shows
performances obtained when the bigram process was further performed
as the preprocessing, and furthermore the number of selectable
utterance sentences in each of the data set for test #1 and the
data set for test #2 was reduced by deleting utterance sentences
having low relevance to a correct utterance sentence from the
selectable utterances.
[0204] As can be seen from FIG. 31, "+replacement" showed higher
performances than "Original". In the case where the data set for
test #2 including unknown words was used, the result showed high
performance similar to that obtained for the data set for test #1
including no unknown words. From these results, it can be seen that
the replacement process according to the present disclosure made it
possible for the dialogue system model generated via learning to
accurately estimate correct responses to utterances of the
user.
[0205] From FIG. 31, it also can be seen that "+replacement,
bi-gram, removing candidate" showed higher performances than
"+replacement" for both the data set for test #1 and the data set
for test #2. This result indicates that performing the
preprocessing in addition to the replacement process according to
the present disclosure made it possible for the dialogue system
model generated via learning to more accurately estimate correct
responses to utterances of the user.
Second Experimental Result
[0206] FIG. 32 shows a second experimental result for a case where
a task similar to the task for the one dialogue shown in the fifth
embodiment was performed. In the task of this experiment, the
dialogue system performs the task to utter to propose restaurant
candidates in the descending order of the evaluation values until
the user accepts one of the proposed restaurant candidates.
[0207] In FIG. 32, "+replacement" shows performances obtained
respectively using the data set for test #1 and the data set for
test #2 subjected to the replacement such that named entities
indicating restaurant names were replaced with abstract expressions
and numeral values were added thereto in the descending order of
the evaluation values corresponding to the named entities. In FIG.
32, "+replacement, normalize" shows performances obtained
respectively using the data set for test #1 and the data set for
test #2 further subjected to the standardization process as the
preprocessing.
[0208] As can be seen from FIG. 32, "+replacement" showed higher
performances than "Original". In the case where the data set for
test #2 including unknown words was used, the result showed high
performance equal to or higher than that obtained for the data set
for test #1 including no unknown words. It can also be seen that
"+replacement, normalize" showed higher performances than
"+replacement". From these results, it can be seen that the
replacement process according to the present disclosure made it
possible for the dialogue system model generated via learning to
accurately estimate correct responses to utterances of the
user.
Third Experimental Result
[0209] FIG. 33 shows a third experimental result for a case where
the dialogue system performed a task to utter to present
information about a phone number, an address, and the like of a
restaurant accepted by a user. In FIG. 33, "+replacement" and
"+replacement, normalize" showed similar results to those shown in
FIG. 32, and thus a further description thereof is omitted.
[0210] In FIG. 33, like FIG. 32, "+replacement" showed higher
performances than "Original". Also in the case where the data set
for test #2 including unknown words was used, the result showed
high performance similar to or higher than the result for the data
set for test #1 including no unknown words. It also can be seen
that "+replacement, normalize" showed higher performances than
"+replacement". These results also show that the replacement
process according to the present disclosure made it possible for
the dialogue system model generated via learning to accurately
estimate correct responses to utterances of the user.
[0211] The information processing apparatus, the information
processing method, and related techniques have been described above
with reference to the specific embodiments. However, the
information processing apparatus, the information processing
method, and related techniques according to the present disclosure
are not limited to these embodiments.
[0212] Each of the processing units in the information processing
apparatus according to any one of the embodiments described above
may be typically implemented by an integrated circuit such as an
LSI. The LSI may be individually formed as chips, or one chip may
be formed so as to include a part or all of the processing
units.
[0213] However, the technique of implementing an integrated circuit
is not limited to the LSI and may be realized by using a dedicated
circuit or a general-purpose processor. An FPGA (Field Programmable
Gate Array) that can be programmed after the manufacture of the LSI
or a reconfigurable processor in which the connections and the
settings of circuit cells disposed inside the LSI can be
reconfigured may be used.
[0214] Note that the present disclosure may also be implemented as
an information processing method executed by an information
processing apparatus.
[0215] In each embodiment described above, each constituent element
may be realized using dedicated hardware or may be realized by
executing software program corresponding to the constituent
element. Each constituent element may be realized by a program
execution unit such as a CPU, a process or the like by reading
software program stored in a storage medium such a hard disk, a
semiconductor memory, or the like and executing the software
program.
[0216] Note that the manners of dividing the meaning generation
apparatus into functional blocks shown in block diagrams are merely
examples. A plurality of functional blocks may be combined together
into one functional block, or one functional block may be divided
into a plurality of functional block. A part of function may be
transferred to another functional block. Functions of a plurality
of functional blocks having similar functions may be executed by
single hardware or may be executed by software in parallel or in a
time-sharing manner.
[0217] The orders of executing the steps described in flow charts
are merely examples that may be employed to realize the present
disclosure, and the order may be different from those employed in
these examples. Part of the steps may be performed concurrently (in
parallel) with other steps.
[0218] The information processing apparatus according to one or
more embodiments has been described above. Note that the present
disclosure is not limited to the embodiments described above. It
will be apparent to those skilled in the art that many various
modifications may be applicable to the embodiments without
departing from the spirit and scope of the present disclosure.
Furthermore, constituent elements of different embodiments may be
combined. In this case, any resultant combination also falls within
the scope of the present disclosure.
[0219] The present disclosure is usable in an information
processing method, an information processing apparatus, and a
program, and more particularly, the present disclosure is usable in
an information processing method, an information processing
apparatus, and a program applied to a task-oriented dialogue system
such as a restaurant reservation system, a hotel reservation
system, a rent apartment search system, and the like.
* * * * *